Research map / 3D vision / Depth and stereo
IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
IRIS recovers physically based materials, spatially-varying HDR lighting, and camera response functions from multi-view LDR images of indoor scenes via alternating optimization with HDR emission restoration.
Problem
Inverse rendering of indoor scenes typically requires HDR input, which limits accessibility; LDR images lose lighting information through clipping and nonlinear CRF mapping.
Approach
Explicitly models LDR image formation (clipping and CRF via EMoR parameterization) inside a factorized physically based inverse rendering pipeline, automatically detects emitters from saturated pixels, and alternates between HDR emission restoration, shading baking, and BRDF/CRF optimization.
Key results
- On FIPT synthetic scenes, IRIS achieves kd PSNR 22.33 and roughness PSNR 21.38 from LDR input, versus 15.49 for FIPT* and 16.85 for NeILF.
- Relighting PSNR on synthetic scenes is 22.861 for IRIS versus 22.444 for Li et al. and 11.626 for FIPT* (LDR inputs).
- Ablation shows removing HDR restoration drops the emission L2 error from 2.35 to 6.23 and roughness PSNR from 21.38 to 7.31 relative to the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add emitter mask Me (ablation) | IRIS · FIPT synthetic scenes | kd PSNR: 16.50 -> 22.33 | Helped |
| add CRF modeling (ablation) | IRIS · FIPT synthetic scenes | kd PSNR: 18.18 -> 22.33 | Helped |
| add HDR restoration (ablation) | IRIS · FIPT synthetic scenes | kd PSNR: 18.20 -> 22.33 | Helped |
| add albedo loss Lalbedo (ablation) | IRIS · FIPT synthetic scenes | kd PSNR: 21.59 -> 22.33 | Helped |
| add CRF loss LCRF (ablation) | IRIS · FIPT synthetic scenes | kd PSNR: 14.49 -> 22.33 | Helped |
| use IRIS instead of FIPT* with LDR input | inverse rendering · FIPT synthetic scenes | kd PSNR: 15.49 -> 22.33 | Helped |
| use IRIS instead of Li et al. | relighting · FIPT synthetic scenes | PSNR: 22.444 -> 22.861 | Helped |
| use IRIS instead of NeILF | novel-view synthesis · FIPT synthetic scenes | PSNR: 30.031 -> 29.468 | No clear effect |
Surprising findings
- FIPT* with the estimated emitter mask still fails to recover emission radiance and BRDF accuracy, showing emitter masks alone are insufficient without HDR restoration.
Limitations
- Relies on an off-the-shelf monocular albedo estimator and BakedSDF geometry for initialization.
- Assumes primary light sources are observed in input images.
- Quantitative evaluation limited to four synthetic scenes; real scenes lack ground truth.
Experimental setup
Datasets: ScanNet++ FIPT synthetic and real scenes
Benchmarks: FIPT synthetic scenes ScanNet++
Metrics: PSNR IoU L2 SSIM LPIPS
Baselines: NeILF FIPT Li et al. single-image inverse rendering
Builds on: FIPT BakedSDF MonoSDF EMoR Cook-Torrance BRDF
Closest papers by meaning
- VMINer: Versatile Multi-view Inverse Rendering with Near- and Far-field Light Sources (CVPR 2024 · similarity 0.88)
- IllumiNeRF: 3D Relighting Without Inverse Rendering (NeurIPS 2024 · similarity 0.86)
- RobIR: Robust Inverse Rendering for High-Illumination Scenes (NeurIPS 2024 · similarity 0.86)
- Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation (CVPR 2025 · similarity 0.86)
- Inverse Rendering using Multi-Bounce Path Tracing and Reservoir Sampling (ICLR 2025 · similarity 0.86)
- SVG-IR: Spatially-Varying Gaussian Splatting for Inverse Rendering (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.